The sys_thread_3() function contains an empty while loop, which by the standard can be optimized away. Please refer to discussion in https://github.com/llvm/llvm-project/issues/60622.
The xlc17 compiler is doing so, and IBM claims that they are following the standard and will not fix this on compiler side.
So we have (at least) 3 ways to circumvent this behavior.

1. We can introduce the call of a nop library function, which will hinder the optimizer to throw away the loop (This is our proposed solution, but instead of a heavy looping thread, the result is a more or less idle thread):
#include <unistd.h>
static void
sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p)
{
                while (1) {
                  sleep(1);
                }
}

2. We can make use of a volatile variable in the loop body which also hinders the optimizer to throw away the loop:
static void
sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p)
{
                volatile int i = 1;
                while (i) {
                  i += 2;
                }
}

3. We can use the __attribute__ ((optnone)) modifier in the function declaration to suppress the optimization at all:
static void
sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p) __attribute__ ((optnone))
{
                while (1) {
                }
}

To make the third approach platform independent, we can implement it in the following way:
In globalDefinitions.hpp
#ifndef OPTNONE
#define OPTNONE
#endif

In globalDefinitions_xlc.hpp
// optnone support
//
// To use if a function should not be optimized
// Usage:
// void* func(size_t size) OPTNONE {...}
#define OPTNONE __attribute__(( optnone))

With this we can change libagentthr001.cpp in a platform independent way to
static void
sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p) OPTNONE
{
     while (1) {
     }
}